Deployments And Environments

Operator validates v2 of a model on /development/predict and clicks 'Promote to Production' in the workspace UI.

Promotion atomically points the production URL at the dev-tested image. In-flight production requests against the old image complete on their replicas; new requests are routed to the new image as replicas come up. Treat promotion as a release event and tag it in the operator's deploy log.

v2 fails in production with elevated 5xx. Operator clicks 'Rollback' to v1 in the deployments UI.

Rollback re-routes production URL traffic to the prior image. v2 replicas drain their in-flight requests and terminate. Confirm rollback by observing replica counts and error rate restore. The rollback action does NOT delete v2 — keep it for forensics.

Operator routes 10% of production traffic to v2 and 90% to v1 for 1 hour, then watches error-rate metrics.

Configure the canary split at the deployment level and watch per-version metrics (error rate, p95 latency, quality regressions) on a fixed observation window. Roll forward (100% v2) or roll back (100% v1) at the window end; do not let canary linger indefinitely.